linux-2.6-microblaze.git
3 years agoMerge tag 'kvm-s390-next-5.13-1' of git://git.kernel.org/pub/scm/linux/kernel/git...
Paolo Bonzini [Thu, 15 Apr 2021 17:02:13 +0000 (13:02 -0400)]
Merge tag 'kvm-s390-next-5.13-1' of git://git./linux/kernel/git/kvms390/linux into HEAD

KVM: s390: Updates for 5.13

- properly handle MVPG in nesting KVM (vsie)
- allow to forward the yield_to hypercall (diagnose 9c)
- fixes

3 years agoMerge branch 'kvm-tdp-fix-rcu' into HEAD
Paolo Bonzini [Fri, 2 Apr 2021 11:25:32 +0000 (07:25 -0400)]
Merge branch 'kvm-tdp-fix-rcu' into HEAD

3 years agoMerge branch 'kvm-tdp-fix-flushes' into HEAD
Paolo Bonzini [Fri, 2 Apr 2021 11:24:54 +0000 (07:24 -0400)]
Merge branch 'kvm-tdp-fix-flushes' into HEAD

3 years agoMerge branch 'kvm-fix-svm-races' into HEAD
Paolo Bonzini [Wed, 31 Mar 2021 11:35:52 +0000 (07:35 -0400)]
Merge branch 'kvm-fix-svm-races' into HEAD

3 years agoKVM: SVM: ensure that EFER.SVME is set when running nested guest or on nested vmexit
Paolo Bonzini [Wed, 31 Mar 2021 10:28:01 +0000 (06:28 -0400)]
KVM: SVM: ensure that EFER.SVME is set when running nested guest or on nested vmexit

Fixing nested_vmcb_check_save to avoid all TOC/TOU races
is a bit harder in released kernels, so do the bare minimum
by avoiding that EFER.SVME is cleared.  This is problematic
because svm_set_efer frees the data structures for nested
virtualization if EFER.SVME is cleared.

Also check that EFER.SVME remains set after a nested vmexit;
clearing it could happen if the bit is zero in the save area
that is passed to KVM_SET_NESTED_STATE (the save area of the
nested state corresponds to the nested hypervisor's state
and is restored on the next nested vmexit).

Cc: stable@vger.kernel.org
Fixes: 2fcf4876ada ("KVM: nSVM: implement on demand allocation of the nested state")
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: SVM: load control fields from VMCB12 before checking them
Paolo Bonzini [Wed, 31 Mar 2021 10:24:43 +0000 (06:24 -0400)]
KVM: SVM: load control fields from VMCB12 before checking them

Avoid races between check and use of the nested VMCB controls.  This
for example ensures that the VMRUN intercept is always reflected to the
nested hypervisor, instead of being processed by the host.  Without this
patch, it is possible to end up with svm->nested.hsave pointing to
the MSR permission bitmap for nested guests.

This bug is CVE-2021-29657.

Reported-by: Felix Wilhelm <fwilhelm@google.com>
Cc: stable@vger.kernel.org
Fixes: 2fcf4876ada ("KVM: nSVM: implement on demand allocation of the nested state")
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Don't allow TDP MMU to yield when recovering NX pages
Sean Christopherson [Thu, 25 Mar 2021 20:01:19 +0000 (13:01 -0700)]
KVM: x86/mmu: Don't allow TDP MMU to yield when recovering NX pages

Prevent the TDP MMU from yielding when zapping a gfn range during NX
page recovery.  If a flush is pending from a previous invocation of the
zapping helper, either in the TDP MMU or the legacy MMU, but the TDP MMU
has not accumulated a flush for the current invocation, then yielding
will release mmu_lock with stale TLB entries.

That being said, this isn't technically a bug fix in the current code, as
the TDP MMU will never yield in this case.  tdp_mmu_iter_cond_resched()
will yield if and only if it has made forward progress, as defined by the
current gfn vs. the last yielded (or starting) gfn.  Because zapping a
single shadow page is guaranteed to (a) find that page and (b) step
sideways at the level of the shadow page, the TDP iter will break its loop
before getting a chance to yield.

But that is all very, very subtle, and will break at the slightest sneeze,
e.g. zapping while holding mmu_lock for read would break as the TDP MMU
wouldn't be guaranteed to see the present shadow page, and thus could step
sideways at a lower level.

Cc: Ben Gardon <bgardon@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210325200119.1359384-4-seanjc@google.com>
[Add lockdep assertion. - Paolo]
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Ensure TLBs are flushed for TDP MMU during NX zapping
Sean Christopherson [Thu, 25 Mar 2021 20:01:18 +0000 (13:01 -0700)]
KVM: x86/mmu: Ensure TLBs are flushed for TDP MMU during NX zapping

Honor the "flush needed" return from kvm_tdp_mmu_zap_gfn_range(), which
does the flush itself if and only if it yields (which it will never do in
this particular scenario), and otherwise expects the caller to do the
flush.  If pages are zapped from the TDP MMU but not the legacy MMU, then
no flush will occur.

Fixes: 29cf0f5007a2 ("kvm: x86/mmu: NX largepage recovery for TDP MMU")
Cc: stable@vger.kernel.org
Cc: Ben Gardon <bgardon@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210325200119.1359384-3-seanjc@google.com>
Reviewed-by: Ben Gardon <bgardon@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Ensure TLBs are flushed when yielding during GFN range zap
Sean Christopherson [Thu, 25 Mar 2021 20:01:17 +0000 (13:01 -0700)]
KVM: x86/mmu: Ensure TLBs are flushed when yielding during GFN range zap

When flushing a range of GFNs across multiple roots, ensure any pending
flush from a previous root is honored before yielding while walking the
tables of the current root.

Note, kvm_tdp_mmu_zap_gfn_range() now intentionally overwrites its local
"flush" with the result to avoid redundant flushes.  zap_gfn_range()
preserves and return the incoming "flush", unless of course the flush was
performed prior to yielding and no new flush was triggered.

Fixes: 1af4a96025b3 ("KVM: x86/mmu: Yield in TDU MMU iter even if no SPTES changed")
Cc: stable@vger.kernel.org
Reviewed-by: Ben Gardon <bgardon@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210325200119.1359384-2-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: s390: VSIE: fix MVPG handling for prefixing and MSO
Claudio Imbrenda [Mon, 22 Mar 2021 14:05:59 +0000 (15:05 +0100)]
KVM: s390: VSIE: fix MVPG handling for prefixing and MSO

Prefixing needs to be applied to the guest real address to translate it
into a guest absolute address.

The value of MSO needs to be added to a guest-absolute address in order to
obtain the host-virtual.

Fixes: bdf7509bbefa ("s390/kvm: VSIE: correctly handle MVPG when in VSIE")
Reported-by: Janosch Frank <frankja@linux.ibm.com>
Signed-off-by: Claudio Imbrenda <imbrenda@linux.ibm.com>
Reviewed-by: David Hildenbrand <david@redhat.com>
Cc: stable@vger.kernel.org
Link: https://lore.kernel.org/r/20210322140559.500716-3-imbrenda@linux.ibm.com
[borntraeger@de.ibm.com simplify mso]
Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
3 years agoKVM: s390: split kvm_s390_real_to_abs
Claudio Imbrenda [Mon, 22 Mar 2021 14:05:58 +0000 (15:05 +0100)]
KVM: s390: split kvm_s390_real_to_abs

A new function _kvm_s390_real_to_abs will apply prefixing to a real address
with a given prefix value.

The old kvm_s390_real_to_abs becomes now a wrapper around the new function.

This is needed to avoid code duplication in vSIE.

Signed-off-by: Claudio Imbrenda <imbrenda@linux.ibm.com>
Reviewed-by: David Hildenbrand <david@redhat.com>
Reviewed-by: Thomas Huth <thuth@redhat.com>
Cc: stable@vger.kernel.org
Link: https://lore.kernel.org/r/20210322140559.500716-2-imbrenda@linux.ibm.com
Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
3 years agoKVM: s390: VSIE: correctly handle MVPG when in VSIE
Claudio Imbrenda [Mon, 1 Feb 2021 20:54:13 +0000 (21:54 +0100)]
KVM: s390: VSIE: correctly handle MVPG when in VSIE

Correctly handle the MVPG instruction when issued by a VSIE guest.

Fixes: a3508fbe9dc6d ("KVM: s390: vsie: initial support for nested virtualization")
Cc: stable@vger.kernel.org # f85f1baaa189: KVM: s390: split kvm_s390_logical_to_effective
Signed-off-by: Claudio Imbrenda <imbrenda@linux.ibm.com>
Acked-by: Janosch Frank <frankja@linux.ibm.com>
Reviewed-by: David Hildenbrand <david@redhat.com>
Acked-by: Christian Borntraeger <borntraeger@de.ibm.com>
Link: https://lore.kernel.org/r/20210302174443.514363-4-imbrenda@linux.ibm.com
[borntraeger@de.ibm.com: apply fixup from Claudio]
Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
3 years agoKVM: s390: extend kvm_s390_shadow_fault to return entry pointer
Claudio Imbrenda [Mon, 1 Feb 2021 16:26:54 +0000 (17:26 +0100)]
KVM: s390: extend kvm_s390_shadow_fault to return entry pointer

Extend kvm_s390_shadow_fault to return the pointer to the valid leaf
DAT table entry, or to the invalid entry.

Also return some flags in the lower bits of the address:
PEI_DAT_PROT: indicates that DAT protection applies because of the
              protection bit in the segment (or, if EDAT, region) tables.
PEI_NOT_PTE: indicates that the address of the DAT table entry returned
             does not refer to a PTE, but to a segment or region table.

Signed-off-by: Claudio Imbrenda <imbrenda@linux.ibm.com>
Cc: stable@vger.kernel.org
Reviewed-by: Janosch Frank <frankja@de.ibm.com>
Reviewed-by: David Hildenbrand <david@redhat.com>
Reviewed-by: Christian Borntraeger <borntraeger@de.ibm.com>
Link: https://lore.kernel.org/r/20210302174443.514363-3-imbrenda@linux.ibm.com
[borntraeger@de.ibm.com: fold in a fix from Claudio]
Signed-off-by: Christian Borntraeger <borntraeger@de.ibm.com>
3 years agoKVM: x86/mmu: Store the address space ID in the TDP iterator
Sean Christopherson [Mon, 15 Mar 2021 23:38:03 +0000 (16:38 -0700)]
KVM: x86/mmu: Store the address space ID in the TDP iterator

Store the address space ID in the TDP iterator so that it can be
retrieved without having to bounce through the root shadow page.  This
streamlines the code and fixes a Sparse warning about not properly using
rcu_dereference() when grabbing the ID from the root on the fly.

Reported-by: kernel test robot <lkp@intel.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Ben Gardon <bgardon@google.com>
Message-Id: <20210315233803.2706477-5-bgardon@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Factor out tdp_iter_return_to_root
Ben Gardon [Mon, 15 Mar 2021 23:38:02 +0000 (16:38 -0700)]
KVM: x86/mmu: Factor out tdp_iter_return_to_root

In tdp_mmu_iter_cond_resched there is a call to tdp_iter_start which
causes the iterator to continue its walk over the paging structure from
the root. This is needed after a yield as paging structure could have
been freed in the interim.

The tdp_iter_start call is not very clear and something of a hack. It
requires exposing tdp_iter fields not used elsewhere in tdp_mmu.c and
the effect is not obvious from the function name. Factor a more aptly
named function out of tdp_iter_start and call it from
tdp_mmu_iter_cond_resched and tdp_iter_start.

No functional change intended.

Signed-off-by: Ben Gardon <bgardon@google.com>
Message-Id: <20210315233803.2706477-4-bgardon@google.com>
Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Fix RCU usage when atomically zapping SPTEs
Ben Gardon [Mon, 15 Mar 2021 23:38:01 +0000 (16:38 -0700)]
KVM: x86/mmu: Fix RCU usage when atomically zapping SPTEs

Fix a missing rcu_dereference in tdp_mmu_zap_spte_atomic.

Reported-by: kernel test robot <lkp@intel.com>
Signed-off-by: Ben Gardon <bgardon@google.com>
Message-Id: <20210315233803.2706477-3-bgardon@google.com>
Reviewed-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Fix RCU usage in handle_removed_tdp_mmu_page
Ben Gardon [Mon, 15 Mar 2021 23:38:00 +0000 (16:38 -0700)]
KVM: x86/mmu: Fix RCU usage in handle_removed_tdp_mmu_page

The pt passed into handle_removed_tdp_mmu_page does not need RCU
protection, as it is not at any risk of being freed by another thread at
that point. However, the implicit cast from tdp_sptep_t to u64 * dropped
the __rcu annotation without a proper rcu_derefrence. Fix this by
passing the pt as a tdp_ptep_t and then rcu_dereferencing it in
the function.

Suggested-by: Sean Christopherson <seanjc@google.com>
Reported-by: kernel test robot <lkp@intel.com>
Signed-off-by: Ben Gardon <bgardon@google.com>
Message-Id: <20210315233803.2706477-2-bgardon@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Mark the PAE roots as decrypted for shadow paging
Sean Christopherson [Tue, 9 Mar 2021 22:42:07 +0000 (14:42 -0800)]
KVM: x86/mmu: Mark the PAE roots as decrypted for shadow paging

Set the PAE roots used as decrypted to play nice with SME when KVM is
using shadow paging.  Explicitly skip setting the C-bit when loading
CR3 for PAE shadow paging, even though it's completely ignored by the
CPU.  The extra documentation is nice to have.

Note, there are several subtleties at play with NPT.  In addition to
legacy shadow paging, the PAE roots are used for SVM's NPT when either
KVM is 32-bit (uses PAE paging) or KVM is 64-bit and shadowing 32-bit
NPT.  However, 32-bit Linux, and thus KVM, doesn't support SME.  And
64-bit KVM can happily set the C-bit in CR3.  This also means that
keeping __sme_set(root) for 32-bit KVM when NPT is enabled is
conceptually wrong, but functionally ok since SME is 64-bit only.
Leave it as is to avoid unnecessary pollution.

Fixes: d0ec49d4de90 ("kvm/x86/svm: Support Secure Memory Encryption within KVM")
Cc: stable@vger.kernel.org
Cc: Brijesh Singh <brijesh.singh@amd.com>
Cc: Tom Lendacky <thomas.lendacky@amd.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210309224207.1218275-5-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Use '0' as the one and only value for an invalid PAE root
Sean Christopherson [Tue, 9 Mar 2021 22:42:06 +0000 (14:42 -0800)]
KVM: x86/mmu: Use '0' as the one and only value for an invalid PAE root

Use '0' to denote an invalid pae_root instead of '0' or INVALID_PAGE.
Unlike root_hpa, the pae_roots hold permission bits and thus are
guaranteed to be non-zero.  Having to deal with both values leads to
bugs, e.g. failing to set back to INVALID_PAGE, warning on the wrong
value, etc...

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210309224207.1218275-4-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: VMX: Track root HPA instead of EPTP for paravirt Hyper-V TLB flush
Sean Christopherson [Fri, 5 Mar 2021 18:31:23 +0000 (10:31 -0800)]
KVM: VMX: Track root HPA instead of EPTP for paravirt Hyper-V TLB flush

Track the address of the top-level EPT struct, a.k.a. the root HPA,
instead of the EPTP itself for Hyper-V's paravirt TLB flush.  The
paravirt API takes only the address, not the full EPTP, and in theory
tracking the EPTP could lead to false negatives, e.g. if the HPA matched
but the attributes in the EPTP do not.  In practice, such a mismatch is
extremely unlikely, if not flat out impossible, given how KVM generates
the EPTP.

Opportunsitically rename the related fields to use the 'root'
nomenclature, and to prefix them with 'hv_' to connect them to Hyper-V's
paravirt TLB flushing.

Reviewed-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210305183123.3978098-12-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: VMX: Skip additional Hyper-V TLB EPTP flushes if one fails
Sean Christopherson [Fri, 5 Mar 2021 18:31:22 +0000 (10:31 -0800)]
KVM: VMX: Skip additional Hyper-V TLB EPTP flushes if one fails

Skip additional EPTP flushes if one fails when processing EPTPs for
Hyper-V's paravirt TLB flushing.  If _any_ flush fails, KVM falls back
to a full global flush, i.e. additional flushes are unnecessary (and
will likely fail anyways).

Continue processing the loop unless a mismatch was already detected,
e.g. to handle the case where the first flush fails and there is a
yet-to-be-detected mismatch.

Reviewed-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210305183123.3978098-11-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: VMX: Define Hyper-V paravirt TLB flush fields iff Hyper-V is enabled
Sean Christopherson [Fri, 5 Mar 2021 18:31:21 +0000 (10:31 -0800)]
KVM: VMX: Define Hyper-V paravirt TLB flush fields iff Hyper-V is enabled

Ifdef away the Hyper-V specific fields in structs kvm_vmx and vcpu_vmx
as each field has only a single reference outside of the struct itself
that isn't already wrapped in ifdeffery (and both are initialization).

vcpu_vmx.ept_pointer in particular should be wrapped as it is valid if
and only if Hyper-v is active, i.e. non-Hyper-V code cannot rely on it
to actually track the current EPTP (without additional code changes).

Reviewed-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210305183123.3978098-10-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: VMX: Explicitly check for hv_remote_flush_tlb when loading pgd
Sean Christopherson [Fri, 5 Mar 2021 18:31:20 +0000 (10:31 -0800)]
KVM: VMX: Explicitly check for hv_remote_flush_tlb when loading pgd

Explicitly check that kvm_x86_ops.tlb_remote_flush() points at Hyper-V's
implementation for PV flushing instead of assuming that a non-NULL
implementation means running on Hyper-V.  Wrap the related logic in
ifdeffery as hv_remote_flush_tlb() is defined iff CONFIG_HYPERV!=n.

Short term, the explicit check makes it more obvious why a non-NULL
tlb_remote_flush() triggers EPTP shenanigans.  Long term, this will
allow TDX to define its own implementation of tlb_remote_flush() without
running afoul of Hyper-V.

Reviewed-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210305183123.3978098-9-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: VMX: Don't invalidate hv_tlb_eptp if the new EPTP matches
Sean Christopherson [Fri, 5 Mar 2021 18:31:19 +0000 (10:31 -0800)]
KVM: VMX: Don't invalidate hv_tlb_eptp if the new EPTP matches

Don't invalidate the common EPTP, and thus trigger rechecking of EPTPs
across all vCPUs, if the new EPTP matches the old/common EPTP.  In all
likelihood this is a meaningless optimization, but there are (uncommon)
scenarios where KVM can reload the same EPTP.

Reviewed-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210305183123.3978098-8-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: VMX: Invalidate hv_tlb_eptp to denote an EPTP mismatch
Sean Christopherson [Fri, 5 Mar 2021 18:31:18 +0000 (10:31 -0800)]
KVM: VMX: Invalidate hv_tlb_eptp to denote an EPTP mismatch

Drop the dedicated 'ept_pointers_match' field in favor of stuffing
'hv_tlb_eptp' with INVALID_PAGE to mark it as invalid, i.e. to denote
that there is at least one EPTP mismatch.  Use a local variable to
track whether or not a mismatch is detected so that hv_tlb_eptp can be
used to skip redundant flushes.

No functional change intended.

Reviewed-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210305183123.3978098-7-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: VMX: Do Hyper-V TLB flush iff vCPU's EPTP hasn't been flushed
Sean Christopherson [Fri, 5 Mar 2021 18:31:17 +0000 (10:31 -0800)]
KVM: VMX: Do Hyper-V TLB flush iff vCPU's EPTP hasn't been flushed

Combine the for-loops for Hyper-V TLB EPTP checking and flushing, and in
doing so skip flushes for vCPUs whose EPTP matches the target EPTP.

Reviewed-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210305183123.3978098-6-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: VMX: Fold Hyper-V EPTP checking into it's only caller
Sean Christopherson [Fri, 5 Mar 2021 18:31:16 +0000 (10:31 -0800)]
KVM: VMX: Fold Hyper-V EPTP checking into it's only caller

Fold check_ept_pointer_match() into hv_remote_flush_tlb_with_range() in
preparation for combining the kvm_for_each_vcpu loops of the ==CHECK and
!=MATCH statements.

No functional change intended.

Reviewed-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210305183123.3978098-5-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: VMX: Stash kvm_vmx in a local variable for Hyper-V paravirt TLB flush
Sean Christopherson [Fri, 5 Mar 2021 18:31:15 +0000 (10:31 -0800)]
KVM: VMX: Stash kvm_vmx in a local variable for Hyper-V paravirt TLB flush

Capture kvm_vmx in a local variable instead of polluting
hv_remote_flush_tlb_with_range() with to_kvm_vmx(kvm).

No functional change intended.

Reviewed-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210305183123.3978098-4-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: VMX: Track common EPTP for Hyper-V's paravirt TLB flush
Sean Christopherson [Fri, 5 Mar 2021 18:31:14 +0000 (10:31 -0800)]
KVM: VMX: Track common EPTP for Hyper-V's paravirt TLB flush

Explicitly track the EPTP that is common to all vCPUs instead of
grabbing vCPU0's EPTP when invoking Hyper-V's paravirt TLB flush.
Tracking the EPTP will allow optimizing the checks when loading a new
EPTP and will also allow dropping ept_pointer_match, e.g. by marking
the common EPTP as invalid.

This also technically fixes a bug where KVM could theoretically flush an
invalid GPA if all vCPUs have an invalid root.  In practice, it's likely
impossible to trigger a remote TLB flush in such a scenario.  In any
case, the superfluous flush is completely benign.

Reviewed-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210305183123.3978098-3-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86: Get active PCID only when writing a CR3 value
Sean Christopherson [Fri, 5 Mar 2021 18:31:13 +0000 (10:31 -0800)]
KVM: x86: Get active PCID only when writing a CR3 value

Retrieve the active PCID only when writing a guest CR3 value, i.e. don't
get the PCID when using EPT or NPT.  The PCID is especially problematic
for EPT as the bits have different meaning, and so the PCID and must be
manually stripped, which is annoying and unnecessary.  And on VMX,
getting the active PCID also involves reading the guest's CR3 and
CR4.PCIDE, i.e. may add pointless VMREADs.

Opportunistically rename the pgd/pgd_level params to root_hpa and
root_level to better reflect their new roles.  Keep the function names,
as "load the guest PGD" is still accurate/correct.

Last, and probably least, pass root_hpa as a hpa_t/u64 instead of an
unsigned long.  The EPTP holds a 64-bit value, even in 32-bit mode, so
in theory EPT could support HIGHMEM for 32-bit KVM.  Never mind that
doing so would require changing the MMU page allocators and reworking
the MMU to use kmap().

Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210305183123.3978098-2-seanjc@google.com>
Reviewed-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM/SVM: Move vmenter.S exception fixups out of line
Uros Bizjak [Fri, 26 Feb 2021 12:56:21 +0000 (13:56 +0100)]
KVM/SVM: Move vmenter.S exception fixups out of line

Avoid jump by moving exception fixups out of line.

Cc: Sean Christopherson <seanjc@google.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Uros Bizjak <ubizjak@gmail.com>
Message-Id: <20210226125621.111723-1-ubizjak@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Dump reserved bits if they're detected on non-MMIO SPTE
Sean Christopherson [Thu, 25 Feb 2021 20:47:49 +0000 (12:47 -0800)]
KVM: x86/mmu: Dump reserved bits if they're detected on non-MMIO SPTE

Debugging unexpected reserved bit page faults sucks.  Dump the reserved
bits that (likely) caused the page fault to make debugging suck a little
less.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210225204749.1512652-25-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Use low available bits for removed SPTEs
Sean Christopherson [Thu, 25 Feb 2021 20:47:48 +0000 (12:47 -0800)]
KVM: x86/mmu: Use low available bits for removed SPTEs

Use low "available" bits to tag REMOVED SPTEs.  Using a high bit is
moderately costly as it often causes the compiler to generate a 64-bit
immediate.  More importantly, this makes it very clear REMOVED_SPTE is
a value, not a flag.

Cc: Ben Gardon <bgardon@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210225204749.1512652-24-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Use is_removed_spte() instead of open coded equivalents
Sean Christopherson [Thu, 25 Feb 2021 20:47:47 +0000 (12:47 -0800)]
KVM: x86/mmu: Use is_removed_spte() instead of open coded equivalents

Use the is_removed_spte() helper instead of open coding the check.

No functional change intended.

Cc: Ben Gardon <bgardon@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210225204749.1512652-23-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Tweak auditing WARN for A/D bits to !PRESENT (was MMIO)
Sean Christopherson [Thu, 25 Feb 2021 20:47:46 +0000 (12:47 -0800)]
KVM: x86/mmu: Tweak auditing WARN for A/D bits to !PRESENT (was MMIO)

Tweak the MMU_WARN that guards against weirdness when querying A/D status
to fire on a !MMU_PRESENT SPTE, as opposed to a MMIO SPTE.  Attempting to
query A/D status on any kind of !MMU_PRESENT SPTE, MMIO or otherwise,
indicates a KVM bug.  Case in point, several now-fixed bugs were
identified by enabling this new WARN.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210225204749.1512652-22-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Use a dedicated bit to track shadow/MMU-present SPTEs
Sean Christopherson [Thu, 25 Feb 2021 20:47:45 +0000 (12:47 -0800)]
KVM: x86/mmu: Use a dedicated bit to track shadow/MMU-present SPTEs

Introduce MMU_PRESENT to explicitly track which SPTEs are "present" from
the MMU's perspective.  Checking for shadow-present SPTEs is a very
common operation for the MMU, particularly in hot paths such as page
faults.  With the addition of "removed" SPTEs for the TDP MMU,
identifying shadow-present SPTEs is quite costly especially since it
requires checking multiple 64-bit values.

On 64-bit KVM, this reduces the footprint of kvm.ko's .text by ~2k bytes.
On 32-bit KVM, this increases the footprint by ~200 bytes, but only
because gcc now inlines several more MMU helpers, e.g. drop_parent_pte().

We now need to drop bit 11, used for the MMU_PRESENT flag, from
the set of bits used to store the generation number in MMIO SPTEs.
Otherwise MMIO SPTEs with bit 11 set would get false positives for
is_shadow_present_spte() and lead to a variety of fireworks, from oopses
to likely hangs of the host kernel.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210225204749.1512652-21-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Use high bits for host/mmu writable masks for EPT SPTEs
Sean Christopherson [Thu, 25 Feb 2021 20:47:44 +0000 (12:47 -0800)]
KVM: x86/mmu: Use high bits for host/mmu writable masks for EPT SPTEs

Use bits 57 and 58 for HOST_WRITABLE and MMU_WRITABLE when using EPT.
This will allow using bit 11 as a constant MMU_PRESENT, which is
desirable as checking for a shadow-present SPTE is one of the most
common SPTE operations in KVM, particular in hot paths such as page
faults.

EPT is short on low available bits; currently only bit 11 is the only
always-available bit.  Bit 10 is also available, but only while KVM
doesn't support mode-based execution.  On the other hand, PAE paging
doesn't have _any_ high available bits.  Thus, using bit 11 is the only
feasible option for MMU_PRESENT.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210225204749.1512652-20-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Make Host-writable and MMU-writable bit locations dynamic
Sean Christopherson [Thu, 25 Feb 2021 20:47:43 +0000 (12:47 -0800)]
KVM: x86/mmu: Make Host-writable and MMU-writable bit locations dynamic

Make the location of the HOST_WRITABLE and MMU_WRITABLE configurable for
a given KVM instance.  This will allow EPT to use high available bits,
which in turn will free up bit 11 for a constant MMU_PRESENT bit.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210225204749.1512652-19-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Move logic for setting SPTE masks for EPT into the MMU proper
Sean Christopherson [Thu, 25 Feb 2021 20:47:42 +0000 (12:47 -0800)]
KVM: x86/mmu: Move logic for setting SPTE masks for EPT into the MMU proper

Let the MMU deal with the SPTE masks to avoid splitting the logic and
knowledge across the MMU and VMX.

The SPTE masks that are used for EPT are very, very tightly coupled to
the MMU implementation.  The use of available bits, the existence of A/D
types, the fact that shadow_x_mask even exists, and so on and so forth
are all baked into the MMU implementation.  Cross referencing the params
to the masks is also a nightmare, as pretty much every param is a u64.

A future patch will make the location of the MMU_WRITABLE and
HOST_WRITABLE bits MMU specific, to free up bit 11 for a MMU_PRESENT bit.
Doing that change with the current kvm_mmu_set_mask_ptes() would be an
absolute mess.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210225204749.1512652-18-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Co-locate code for setting various SPTE masks
Sean Christopherson [Thu, 25 Feb 2021 20:47:41 +0000 (12:47 -0800)]
KVM: x86/mmu: Co-locate code for setting various SPTE masks

Squish all the code for (re)setting the various SPTE masks into one
location.  With the split code, it's not at all clear that the masks are
set once during module initialization.  This will allow a future patch to
clean up initialization of the masks without shuffling code all over
tarnation.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210225204749.1512652-17-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Move initial kvm_mmu_set_mask_ptes() call into MMU proper
Sean Christopherson [Thu, 25 Feb 2021 20:47:40 +0000 (12:47 -0800)]
KVM: x86/mmu: Move initial kvm_mmu_set_mask_ptes() call into MMU proper

Move kvm_mmu_set_mask_ptes() into mmu.c as prep for future cleanup of the
mask initialization code.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210225204749.1512652-16-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Document dependency bewteen TDP A/D type and saved bits
Sean Christopherson [Thu, 25 Feb 2021 20:47:39 +0000 (12:47 -0800)]
KVM: x86/mmu: Document dependency bewteen TDP A/D type and saved bits

Document that SHADOW_ACC_TRACK_SAVED_BITS_SHIFT is directly dependent on
bits 53:52 being used to track the A/D type.

Remove PT64_SECOND_AVAIL_BITS_SHIFT as it is at best misleading, and at
worst wrong.  For PAE paging, which arguably is a variant of PT64, the
bits are reserved.  For MMIO SPTEs the bits are not available as they're
used for the MMIO generation.  For access tracked SPTEs, they are also
not available as bits 56:54 are used to store the original RX bits.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210225204749.1512652-15-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Use MMIO SPTE bits 53 and 52 for the MMIO generation
Sean Christopherson [Thu, 25 Feb 2021 20:47:38 +0000 (12:47 -0800)]
KVM: x86/mmu: Use MMIO SPTE bits 53 and 52 for the MMIO generation

Use bits 53 and 52 for the MMIO generation now that they're not used to
identify MMIO SPTEs.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210225204749.1512652-14-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Rename and document A/D scheme for TDP SPTEs
Sean Christopherson [Thu, 25 Feb 2021 20:47:37 +0000 (12:47 -0800)]
KVM: x86/mmu: Rename and document A/D scheme for TDP SPTEs

Rename the various A/D status defines to explicitly associated them with
TDP.  There is a subtle dependency on the bits in question never being
set when using PAE paging, as those bits are reserved, not available.
I.e. using these bits outside of TDP (technically EPT) would cause
explosions.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210225204749.1512652-13-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Add module param to disable MMIO caching (for testing)
Sean Christopherson [Thu, 25 Feb 2021 20:47:36 +0000 (12:47 -0800)]
KVM: x86/mmu: Add module param to disable MMIO caching (for testing)

Add a module param to disable MMIO caching so that it's possible to test
the related flows without access to the necessary hardware.  Using shadow
paging with 64-bit KVM and 52 bits of physical address space must disable
MMIO caching as there are no reserved bits to be had.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210225204749.1512652-12-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Stop using software available bits to denote MMIO SPTEs
Sean Christopherson [Thu, 25 Feb 2021 20:47:35 +0000 (12:47 -0800)]
KVM: x86/mmu: Stop using software available bits to denote MMIO SPTEs

Stop tagging MMIO SPTEs with specific available bits and instead detect
MMIO SPTEs by checking for their unique SPTE value.  The value is
guaranteed to be unique on shadow paging and NPT as setting reserved
physical address bits on any other type of SPTE would consistute a KVM
bug.  Ditto for EPT, as creating a WX non-MMIO would also be a bug.

Note, this approach is also future-compatibile with TDX, which will need
to reflect MMIO EPT violations as #VEs into the guest.  To create an EPT
violation instead of a misconfig, TDX EPTs will need to have RWX=0,  But,
MMIO SPTEs will also be the only case where KVM clears SUPPRESS_VE, so
MMIO SPTEs will still be guaranteed to have a unique value within a given
MMU context.

The main motivation is to make it easier to reason about which types of
SPTEs use which available bits.  As a happy side effect, this frees up
two more bits for storing the MMIO generation.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210225204749.1512652-11-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Rename 'mask' to 'spte' in MMIO SPTE helpers
Sean Christopherson [Thu, 25 Feb 2021 20:47:34 +0000 (12:47 -0800)]
KVM: x86/mmu: Rename 'mask' to 'spte' in MMIO SPTE helpers

The value returned by make_mmio_spte() is a SPTE, it is not a mask.
Name it accordingly.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210225204749.1512652-10-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Drop redundant trace_kvm_mmu_set_spte() in the TDP MMU
Sean Christopherson [Thu, 25 Feb 2021 20:47:33 +0000 (12:47 -0800)]
KVM: x86/mmu: Drop redundant trace_kvm_mmu_set_spte() in the TDP MMU

Remove TDP MMU's call to trace_kvm_mmu_set_spte() that is done for both
shadow-present SPTEs and MMIO SPTEs.  It's fully redundant for the
former, and unnecessary for the latter.  This aligns TDP MMU tracing
behavior with that of the legacy MMU.

Fixes: 33dd3574f5fe ("kvm: x86/mmu: Add existing trace points to TDP MMU")
Cc: Ben Gardon <bgardon@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210225204749.1512652-9-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Handle MMIO SPTEs directly in mmu_set_spte()
Sean Christopherson [Thu, 25 Feb 2021 20:47:32 +0000 (12:47 -0800)]
KVM: x86/mmu: Handle MMIO SPTEs directly in mmu_set_spte()

Now that it should be impossible to convert a valid SPTE to an MMIO SPTE,
handle MMIO SPTEs early in mmu_set_spte() without going through
set_spte() and all the logic for removing an existing, valid SPTE.
The other caller of set_spte(), FNAME(sync_page)(), explicitly handles
MMIO SPTEs prior to calling set_spte().

This simplifies mmu_set_spte() and set_spte(), and also "fixes" an oddity
where MMIO SPTEs are traced by both trace_kvm_mmu_set_spte() and
trace_mark_mmio_spte().

Note, mmu_spte_set() will WARN if this new approach causes KVM to create
an MMIO SPTE overtop a valid SPTE.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210225204749.1512652-8-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Don't install bogus MMIO SPTEs if MMIO caching is disabled
Sean Christopherson [Thu, 25 Feb 2021 20:47:31 +0000 (12:47 -0800)]
KVM: x86/mmu: Don't install bogus MMIO SPTEs if MMIO caching is disabled

If MMIO caching is disabled, e.g. when using shadow paging on CPUs with
52 bits of PA space, go straight to MMIO emulation and don't install an
MMIO SPTE.  The SPTE will just generate a !PRESENT #PF, i.e. can't
actually accelerate future MMIO.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210225204749.1512652-7-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Retry page faults that hit an invalid memslot
Sean Christopherson [Thu, 25 Feb 2021 20:47:30 +0000 (12:47 -0800)]
KVM: x86/mmu: Retry page faults that hit an invalid memslot

Retry page faults (re-enter the guest) that hit an invalid memslot
instead of treating the memslot as not existing, i.e. handling the
page fault as an MMIO access.  When deleting a memslot, SPTEs aren't
zapped and the TLBs aren't flushed until after the memslot has been
marked invalid.

Handling the invalid slot as MMIO means there's a small window where a
page fault could replace a valid SPTE with an MMIO SPTE.  The legacy
MMU handles such a scenario cleanly, but the TDP MMU assumes such
behavior is impossible (see the BUG() in __handle_changed_spte()).
There's really no good reason why the legacy MMU should allow such a
scenario, and closing this hole allows for additional cleanups.

Fixes: 2f2fad0897cb ("kvm: x86/mmu: Add functions to handle changed TDP SPTEs")
Cc: Ben Gardon <bgardon@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210225204749.1512652-6-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Disable MMIO caching if MMIO value collides with L1TF
Sean Christopherson [Thu, 25 Feb 2021 20:47:29 +0000 (12:47 -0800)]
KVM: x86/mmu: Disable MMIO caching if MMIO value collides with L1TF

Disable MMIO caching if the MMIO value collides with the L1TF mitigation
that usurps high PFN bits.  In practice this should never happen as only
CPUs with SME support can generate such a collision (because the MMIO
value can theoretically get adjusted into legal memory), and no CPUs
exist that support SME and are susceptible to L1TF.  But, closing the
hole is trivial.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210225204749.1512652-5-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Bail from fast_page_fault() if SPTE is not shadow-present
Sean Christopherson [Thu, 25 Feb 2021 20:47:28 +0000 (12:47 -0800)]
KVM: x86/mmu: Bail from fast_page_fault() if SPTE is not shadow-present

Bail from fast_page_fault() if the SPTE is not a shadow-present SPTE.
Functionally, this is not strictly necessary as the !is_access_allowed()
check will eventually reject the fast path, but an early check on
shadow-present skips unnecessary checks and will allow a future patch to
tweak the A/D status auditing to warn if KVM attempts to query A/D bits
without first ensuring the SPTE is a shadow-present SPTE.

Note, is_shadow_present_pte() is quite expensive at this time, i.e. this
might be a net negative in the short term.  A future patch will optimize
is_shadow_present_pte() to a single AND operation and remedy the issue.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210225204749.1512652-4-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Check for shadow-present SPTE before querying A/D status
Sean Christopherson [Thu, 25 Feb 2021 20:47:27 +0000 (12:47 -0800)]
KVM: x86/mmu: Check for shadow-present SPTE before querying A/D status

When updating accessed and dirty bits, check that the new SPTE is present
before attempting to query its A/D bits.  Failure to confirm the SPTE is
present can theoretically cause a false negative, e.g. if a MMIO SPTE
replaces a "real" SPTE and somehow the PFNs magically match.

Realistically, this is all but guaranteed to be a benign bug.  Fix it up
primarily so that a future patch can tweak the MMU_WARN_ON checking A/D
status to fire if the SPTE is not-present.

Fixes: f8e144971c68 ("kvm: x86/mmu: Add access tracking for tdp_mmu")
Cc: Ben Gardon <bgardon@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210225204749.1512652-3-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Add convenience wrapper for acting on single hva in TDP MMU
Sean Christopherson [Fri, 26 Feb 2021 01:03:29 +0000 (17:03 -0800)]
KVM: x86/mmu: Add convenience wrapper for acting on single hva in TDP MMU

Add a TDP MMU helper to handle a single HVA hook, the name is a nice
reminder that the flow in question is operating on a single HVA.

No functional change intended.

Cc: Ben Gardon <bgardon@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210226010329.1766033-6-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Add typedefs for rmap/iter handlers
Sean Christopherson [Fri, 26 Feb 2021 01:03:28 +0000 (17:03 -0800)]
KVM: x86/mmu: Add typedefs for rmap/iter handlers

Add typedefs for the MMU handlers that are invoked when walking the MMU
SPTEs (rmaps in legacy MMU) to act on a host virtual address range.

No functional change intended.

Cc: Ben Gardon <bgardon@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210226010329.1766033-5-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Use 'end' param in TDP MMU's test_age_gfn()
Sean Christopherson [Fri, 26 Feb 2021 01:03:27 +0000 (17:03 -0800)]
KVM: x86/mmu: Use 'end' param in TDP MMU's test_age_gfn()

Use the @end param when aging a GFN instead of hardcoding the walk to a
single GFN.  Unlike tdp_set_spte(), which simply cannot work with more
than one GFN, aging multiple GFNs would not break, though admittedly it
would be weird.  Be nice to the casual reader and don't make them puzzle
out why the end GFN is unused.

No functional change intended.

Cc: Ben Gardon <bgardon@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210226010329.1766033-4-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: WARN if TDP MMU's set_tdp_spte() sees multiple GFNs
Sean Christopherson [Fri, 26 Feb 2021 01:03:26 +0000 (17:03 -0800)]
KVM: x86/mmu: WARN if TDP MMU's set_tdp_spte() sees multiple GFNs

WARN if set_tdp_spte() is invoked with multipel GFNs.  It is specifically
a callback to handle a single host PTE being changed.  Consuming the
@end parameter also eliminates the confusing 'unused' parameter.

Cc: Ben Gardon <bgardon@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210226010329.1766033-3-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Remove spurious TLB flush from TDP MMU's change_pte() hook
Sean Christopherson [Fri, 26 Feb 2021 01:03:25 +0000 (17:03 -0800)]
KVM: x86/mmu: Remove spurious TLB flush from TDP MMU's change_pte() hook

Remove an unnecessary remote TLB flush from set_tdp_spte(), the TDP MMu's
hook for handling change_pte() invocations from the MMU notifier.  If
the new host PTE is writable, the flush is completely redundant as there
are no futher changes to the SPTE before the post-loop flush.  If the
host PTE is read-only, then the primary MMU is responsible for ensuring
that the contents of the old and new pages are identical, thus it's safe
to let the guest continue reading both the old and new pages.  KVM must
only ensure the old page cannot be referenced after returning from its
callback; this is handled by the post-loop flush.

Fixes: 1d8dd6b3f12b ("kvm: x86/mmu: Support changed pte notifier in tdp MMU")
Cc: Ben Gardon <bgardon@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210226010329.1766033-2-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86: mmu: initialize fault.async_page_fault in walk_addr_generic
Maxim Levitsky [Thu, 25 Feb 2021 15:41:33 +0000 (17:41 +0200)]
KVM: x86: mmu: initialize fault.async_page_fault in walk_addr_generic

This field was left uninitialized by a mistake.

Signed-off-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20210225154135.405125-3-mlevitsk@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86: determine if an exception has an error code only when injecting it.
Maxim Levitsky [Thu, 25 Feb 2021 15:41:32 +0000 (17:41 +0200)]
KVM: x86: determine if an exception has an error code only when injecting it.

A page fault can be queued while vCPU is in real paged mode on AMD, and
AMD manual asks the user to always intercept it
(otherwise result is undefined).
The resulting VM exit, does have an error code.

Signed-off-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20210225154135.405125-2-mlevitsk@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: nSVM: Optimize vmcb12 to vmcb02 save area copies
Cathy Avery [Mon, 1 Mar 2021 20:08:44 +0000 (15:08 -0500)]
KVM: nSVM: Optimize vmcb12 to vmcb02 save area copies

Use the vmcb12 control clean field to determine which vmcb12.save
registers were marked dirty in order to minimize register copies
when switching from L1 to L2. Those vmcb12 registers marked as dirty need
to be copied to L0's vmcb02 as they will be used to update the vmcb
state cache for the L2 VMRUN.  In the case where we have a different
vmcb12 from the last L2 VMRUN all vmcb12.save registers must be
copied over to L2.save.

Tested:
kvm-unit-tests
kvm selftests
Fedora L1 L2

Suggested-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Cathy Avery <cavery@redhat.com>
Message-Id: <20210301200844.2000-1-cavery@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: SVM: Add support for Virtual SPEC_CTRL
Babu Moger [Wed, 17 Feb 2021 15:56:04 +0000 (10:56 -0500)]
KVM: SVM: Add support for Virtual SPEC_CTRL

Newer AMD processors have a feature to virtualize the use of the
SPEC_CTRL MSR. Presence of this feature is indicated via CPUID
function 0x8000000A_EDX[20]: GuestSpecCtrl. Hypervisors are not
required to enable this feature since it is automatically enabled on
processors that support it.

A hypervisor may wish to impose speculation controls on guest
execution or a guest may want to impose its own speculation controls.
Therefore, the processor implements both host and guest
versions of SPEC_CTRL.

When in host mode, the host SPEC_CTRL value is in effect and writes
update only the host version of SPEC_CTRL. On a VMRUN, the processor
loads the guest version of SPEC_CTRL from the VMCB. When the guest
writes SPEC_CTRL, only the guest version is updated. On a VMEXIT,
the guest version is saved into the VMCB and the processor returns
to only using the host SPEC_CTRL for speculation control. The guest
SPEC_CTRL is located at offset 0x2E0 in the VMCB.

The effective SPEC_CTRL setting is the guest SPEC_CTRL setting or'ed
with the hypervisor SPEC_CTRL setting. This allows the hypervisor to
ensure a minimum SPEC_CTRL if desired.

This support also fixes an issue where a guest may sometimes see an
inconsistent value for the SPEC_CTRL MSR on processors that support
this feature. With the current SPEC_CTRL support, the first write to
SPEC_CTRL is intercepted and the virtualized version of the SPEC_CTRL
MSR is not updated. When the guest reads back the SPEC_CTRL MSR, it
will be 0x0, instead of the actual expected value. There isn’t a
security concern here, because the host SPEC_CTRL value is or’ed with
the Guest SPEC_CTRL value to generate the effective SPEC_CTRL value.
KVM writes with the guest's virtualized SPEC_CTRL value to SPEC_CTRL
MSR just before the VMRUN, so it will always have the actual value
even though it doesn’t appear that way in the guest. The guest will
only see the proper value for the SPEC_CTRL register if the guest was
to write to the SPEC_CTRL register again. With Virtual SPEC_CTRL
support, the save area spec_ctrl is properly saved and restored.
So, the guest will always see the proper value when it is read back.

Signed-off-by: Babu Moger <babu.moger@amd.com>
Message-Id: <161188100955.28787.11816849358413330720.stgit@bmoger-ubuntu>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agox86/cpufeatures: Add the Virtual SPEC_CTRL feature
Babu Moger [Fri, 29 Jan 2021 00:43:22 +0000 (18:43 -0600)]
x86/cpufeatures: Add the Virtual SPEC_CTRL feature

Newer AMD processors have a feature to virtualize the use of the
SPEC_CTRL MSR. Presence of this feature is indicated via CPUID
function 0x8000000A_EDX[20]: GuestSpecCtrl. When present, the
SPEC_CTRL MSR is automatically virtualized.

Signed-off-by: Babu Moger <babu.moger@amd.com>
Acked-by: Borislav Petkov <bp@suse.de>
Message-Id: <161188100272.28787.4097272856384825024.stgit@bmoger-ubuntu>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: nSVM: always use vmcb01 to for vmsave/vmload of guest state
Maxim Levitsky [Wed, 10 Feb 2021 16:54:36 +0000 (18:54 +0200)]
KVM: nSVM: always use vmcb01 to for vmsave/vmload of guest state

This allows to avoid copying of these fields between vmcb01
and vmcb02 on nested guest entry/exit.

Signed-off-by: Maxim Levitsky <mlevitsk@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: SVM: move VMLOAD/VMSAVE to C code
Paolo Bonzini [Tue, 9 Feb 2021 12:10:43 +0000 (07:10 -0500)]
KVM: SVM: move VMLOAD/VMSAVE to C code

Thanks to the new macros that handle exception handling for SVM
instructions, it is easier to just do the VMLOAD/VMSAVE in C.
This is safe, as shown by the fact that the host reload is
already done outside the assembly source.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: SVM: Skip intercepted PAUSE instructions after emulation
Sean Christopherson [Fri, 5 Feb 2021 00:57:50 +0000 (16:57 -0800)]
KVM: SVM: Skip intercepted PAUSE instructions after emulation

Skip PAUSE after interception to avoid unnecessarily re-executing the
instruction in the guest, e.g. after regaining control post-yield.
This is a benign bug as KVM disables PAUSE interception if filtering is
off, including the case where pause_filter_count is set to zero.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210205005750.3841462-10-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: SVM: Don't manually emulate RDPMC if nrips=0
Sean Christopherson [Fri, 5 Feb 2021 00:57:49 +0000 (16:57 -0800)]
KVM: SVM: Don't manually emulate RDPMC if nrips=0

Remove bizarre code that causes KVM to run RDPMC through the emulator
when nrips is disabled.  Accelerated emulation of RDPMC doesn't rely on
any additional data from the VMCB, and SVM has generic handling for
updating RIP to skip instructions when nrips is disabled.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210205005750.3841462-9-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86: Move RDPMC emulation to common code
Sean Christopherson [Fri, 5 Feb 2021 00:57:48 +0000 (16:57 -0800)]
KVM: x86: Move RDPMC emulation to common code

Move the entirety of the accelerated RDPMC emulation to x86.c, and assign
the common handler directly to the exit handler array for VMX.  SVM has
bizarre nrips behavior that prevents it from directly invoking the common
handler.  The nrips goofiness will be addressed in a future patch.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210205005750.3841462-8-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86: Move trivial instruction-based exit handlers to common code
Sean Christopherson [Fri, 5 Feb 2021 00:57:47 +0000 (16:57 -0800)]
KVM: x86: Move trivial instruction-based exit handlers to common code

Move the trivial exit handlers, e.g. for instructions that KVM
"emulates" as nops, to common x86 code.  Assign the common handlers
directly to the exit handler arrays and drop the vendor trampolines.

Opportunistically use pr_warn_once() where appropriate.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210205005750.3841462-7-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86: Move XSETBV emulation to common code
Sean Christopherson [Fri, 5 Feb 2021 00:57:46 +0000 (16:57 -0800)]
KVM: x86: Move XSETBV emulation to common code

Move the entirety of XSETBV emulation to x86.c, and assign the
function directly to both VMX's and SVM's exit handlers, i.e. drop the
unnecessary trampolines.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210205005750.3841462-6-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: nSVM: Add VMLOAD/VMSAVE helper to deduplicate code
Sean Christopherson [Fri, 5 Feb 2021 00:57:45 +0000 (16:57 -0800)]
KVM: nSVM: Add VMLOAD/VMSAVE helper to deduplicate code

Add another helper layer for VMLOAD+VMSAVE, the code is identical except
for the one line that determines which VMCB is the source and which is
the destination.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210205005750.3841462-5-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: nSVM: Add helper to synthesize nested VM-Exit without collateral
Sean Christopherson [Tue, 2 Mar 2021 17:45:15 +0000 (09:45 -0800)]
KVM: nSVM: Add helper to synthesize nested VM-Exit without collateral

Add a helper to consolidate boilerplate for nested VM-Exits that don't
provide any data in exit_info_*.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210302174515.2812275-3-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86: Handle triple fault in L2 without killing L1
Sean Christopherson [Tue, 2 Mar 2021 17:45:14 +0000 (09:45 -0800)]
KVM: x86: Handle triple fault in L2 without killing L1

Synthesize a nested VM-Exit if L2 triggers an emulated triple fault
instead of exiting to userspace, which likely will kill L1.  Any flow
that does KVM_REQ_TRIPLE_FAULT is suspect, but the most common scenario
for L2 killing L1 is if L0 (KVM) intercepts a contributory exception that
is _not_intercepted by L1.  E.g. if KVM is intercepting #GPs for the
VMware backdoor, a #GP that occurs in L2 while vectoring an injected #DF
will cause KVM to emulate triple fault.

Cc: Boris Ostrovsky <boris.ostrovsky@oracle.com>
Cc: Jim Mattson <jmattson@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210302174515.2812275-2-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: SVM: Pass struct kvm_vcpu to exit handlers (and many, many other places)
Paolo Bonzini [Tue, 2 Mar 2021 19:40:39 +0000 (14:40 -0500)]
KVM: SVM: Pass struct kvm_vcpu to exit handlers (and many, many other places)

Refactor the svm_exit_handlers API to pass @vcpu instead of @svm to
allow directly invoking common x86 exit handlers (in a future patch).
Opportunistically convert an absurd number of instances of 'svm->vcpu'
to direct uses of 'vcpu' to avoid pointless casting.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210205005750.3841462-4-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: SVM: merge update_cr0_intercept into svm_set_cr0
Paolo Bonzini [Thu, 18 Feb 2021 14:51:06 +0000 (09:51 -0500)]
KVM: SVM: merge update_cr0_intercept into svm_set_cr0

The logic of update_cr0_intercept is pointlessly complicated.
All svm_set_cr0 is compute the effective cr0 and compare it with
the guest value.

Inlining the function and simplifying the condition
clarifies what it is doing.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: nSVM: Trace VM-Enter consistency check failures
Sean Christopherson [Thu, 4 Feb 2021 00:01:17 +0000 (16:01 -0800)]
KVM: nSVM: Trace VM-Enter consistency check failures

Use trace_kvm_nested_vmenter_failed() and its macro magic to trace
consistency check failures on nested VMRUN.  Tracing such failures by
running the buggy VMM as a KVM guest is often the only way to get a
precise explanation of why VMRUN failed.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210204000117.3303214-13-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86: Move nVMX's consistency check macro to common code
Sean Christopherson [Thu, 4 Feb 2021 00:01:16 +0000 (16:01 -0800)]
KVM: x86: Move nVMX's consistency check macro to common code

Move KVM's CC() macro to x86.h so that it can be reused by nSVM.
Debugging VM-Enter is as painful on SVM as it is on VMX.

Rename the more visible macro to KVM_NESTED_VMENTER_CONSISTENCY_CHECK
to avoid any collisions with the uber-concise "CC".

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210204000117.3303214-12-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: nSVM: Add missing checks for reserved bits to svm_set_nested_state()
Krish Sadhukhan [Tue, 6 Oct 2020 19:06:52 +0000 (19:06 +0000)]
KVM: nSVM: Add missing checks for reserved bits to svm_set_nested_state()

The path for SVM_SET_NESTED_STATE needs to have the same checks for the CPU
registers, as we have in the VMRUN path for a nested guest. This patch adds
those missing checks to svm_set_nested_state().

Suggested-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Krish Sadhukhan <krish.sadhukhan@oracle.com>
Message-Id: <20201006190654.32305-3-krish.sadhukhan@oracle.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: nSVM: only copy L1 non-VMLOAD/VMSAVE data in svm_set_nested_state()
Paolo Bonzini [Tue, 17 Nov 2020 07:51:35 +0000 (02:51 -0500)]
KVM: nSVM: only copy L1 non-VMLOAD/VMSAVE data in svm_set_nested_state()

The VMLOAD/VMSAVE data is not taken from userspace, since it will
not be restored on VMEXIT (it will be copied from VMCB02 to VMCB01).
For clarity, replace the wholesale copy of the VMCB save area
with a copy of that state only.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: nSVM: do not mark all VMCB02 fields dirty on nested vmexit
Paolo Bonzini [Mon, 16 Nov 2020 11:38:19 +0000 (06:38 -0500)]
KVM: nSVM: do not mark all VMCB02 fields dirty on nested vmexit

Since L1 and L2 now use different VMCBs, most of the fields remain the
same in VMCB02 from one L2 run to the next.  Since KVM itself is not
looking at VMCB12's clean field, for now not much can be optimized.
However, in the future we could avoid more copies if the VMCB12's SEG
and DT sections are clean.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: nSVM: do not mark all VMCB01 fields dirty on nested vmexit
Paolo Bonzini [Mon, 16 Nov 2020 11:38:19 +0000 (06:38 -0500)]
KVM: nSVM: do not mark all VMCB01 fields dirty on nested vmexit

Since L1 and L2 now use different VMCBs, most of the fields remain
the same from one L1 run to the next.  svm_set_cr0 and other functions
called by nested_svm_vmexit already take care of clearing the
corresponding clean bits; only the TSC offset is special.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: nSVM: do not copy vmcb01->control blindly to vmcb02->control
Paolo Bonzini [Mon, 16 Nov 2020 11:13:15 +0000 (06:13 -0500)]
KVM: nSVM: do not copy vmcb01->control blindly to vmcb02->control

Most fields were going to be overwritten by vmcb12 control fields, or
do not matter at all because they are filled by the processor on vmexit.
Therefore, we need not copy them from vmcb01 to vmcb02 on vmentry.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: nSVM: rename functions and variables according to vmcbXY nomenclature
Paolo Bonzini [Tue, 17 Nov 2020 10:15:41 +0000 (05:15 -0500)]
KVM: nSVM: rename functions and variables according to vmcbXY nomenclature

Now that SVM is using a separate vmcb01 and vmcb02 (and also uses the vmcb12
naming) we can give clearer names to functions that write to and read
from those VMCBs.  Likewise, variables and parameters can be renamed
from nested_vmcb to vmcb12.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: nSVM: Track the ASID generation of the vmcb vmrun through the vmcb
Cathy Avery [Tue, 12 Jan 2021 16:43:13 +0000 (11:43 -0500)]
KVM: nSVM: Track the ASID generation of the vmcb vmrun through the vmcb

This patch moves the asid_generation from the vcpu to the vmcb
in order to track the ASID generation that was active the last
time the vmcb was run. If sd->asid_generation changes between
two runs, the old ASID is invalid and must be changed.

Suggested-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Cathy Avery <cavery@redhat.com>
Message-Id: <20210112164313.4204-3-cavery@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: nSVM: Track the physical cpu of the vmcb vmrun through the vmcb
Cathy Avery [Tue, 12 Jan 2021 16:43:12 +0000 (11:43 -0500)]
KVM: nSVM: Track the physical cpu of the vmcb vmrun through the vmcb

This patch moves the physical cpu tracking from the vcpu
to the vmcb in svm_switch_vmcb. If either vmcb01 or vmcb02
change physical cpus from one vmrun to the next the vmcb's
previous cpu is preserved for comparison with the current
cpu and the vmcb is marked dirty if different. This prevents
the processor from using old cached data for a vmcb that may
have been updated on a prior run on a different processor.

It also moves the physical cpu check from svm_vcpu_load
to pre_svm_run as the check only needs to be done at run.

Suggested-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Cathy Avery <cavery@redhat.com>
Message-Id: <20210112164313.4204-2-cavery@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: SVM: Use a separate vmcb for the nested L2 guest
Cathy Avery [Wed, 13 Jan 2021 12:07:52 +0000 (07:07 -0500)]
KVM: SVM: Use a separate vmcb for the nested L2 guest

svm->vmcb will now point to a separate vmcb for L1 (not nested) or L2
(nested).

The main advantages are removing get_host_vmcb and hsave, in favor of
concepts that are shared with VMX.

We don't need anymore to stash the L1 registers in hsave while L2
runs, but we need to copy the VMLOAD/VMSAVE registers from VMCB01 to
VMCB02 and back.  This more or less has the same cost, but code-wise
nested_svm_vmloadsave can be reused.

This patch omits several optimizations that are possible:

- for simplicity there is some wholesale copying of vmcb.control areas
which can go away.

- we should be able to better use the VMCB01 and VMCB02 clean bits.

- another possibility is to always use VMCB01 for VMLOAD and VMSAVE,
thus avoiding the copy of VMLOAD/VMSAVE registers from VMCB01 to
VMCB02 and back.

Tested:
kvm-unit-tests
kvm self tests
Loaded fedora nested guest on fedora

Signed-off-by: Cathy Avery <cavery@redhat.com>
Message-Id: <20201011184818.3609-3-cavery@redhat.com>
[Fix conflicts; keep VMCB02 G_PAT up to date whenever guest writes the
 PAT MSR; do not copy CR4 over from VMCB01 as it is not needed anymore; add
 a few more comments. - Paolo]
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: nSVM: Set the shadow root level to the TDP level for nested NPT
Sean Christopherson [Fri, 5 Mar 2021 01:10:45 +0000 (17:10 -0800)]
KVM: nSVM: Set the shadow root level to the TDP level for nested NPT

Override the shadow root level in the MMU context when configuring
NPT for shadowing nested NPT.  The level is always tied to the TDP level
of the host, not whatever level the guest happens to be using.

Fixes: 096586fda522 ("KVM: nSVM: Correctly set the shadow NPT root level in its MMU role")
Cc: stable@vger.kernel.org
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210305011101.3597423-2-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: SVM: Don't strip the C-bit from CR2 on #PF interception
Sean Christopherson [Fri, 5 Mar 2021 01:10:56 +0000 (17:10 -0800)]
KVM: SVM: Don't strip the C-bit from CR2 on #PF interception

Don't strip the C-bit from the faulting address on an intercepted #PF,
the address is a virtual address, not a physical address.

Fixes: 0ede79e13224 ("KVM: SVM: Clear C-bit from the page fault address")
Cc: stable@vger.kernel.org
Cc: Brijesh Singh <brijesh.singh@amd.com>
Cc: Tom Lendacky <thomas.lendacky@amd.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210305011101.3597423-13-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: WARN on NULL pae_root or lm_root, or bad shadow root level
Sean Christopherson [Fri, 5 Mar 2021 01:11:01 +0000 (17:11 -0800)]
KVM: x86/mmu: WARN on NULL pae_root or lm_root, or bad shadow root level

WARN if KVM is about to dereference a NULL pae_root or lm_root when
loading an MMU, and convert the BUG() on a bad shadow_root_level into a
WARN (now that errors are handled cleanly).  With nested NPT, botching
the level and sending KVM down the wrong path is all too easy, and the
on-demand allocation of pae_root and lm_root means bugs crash the host.
Obviously, KVM could unconditionally allocate the roots, but that's
arguably a worse failure mode as it would potentially corrupt the guest
instead of crashing it.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210305011101.3597423-18-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Sync roots after MMU load iff load as successful
Sean Christopherson [Fri, 5 Mar 2021 01:11:00 +0000 (17:11 -0800)]
KVM: x86/mmu: Sync roots after MMU load iff load as successful

For clarity, explicitly skip syncing roots if the MMU load failed
instead of relying on the !VALID_PAGE check in kvm_mmu_sync_roots().

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210305011101.3597423-17-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Unexport MMU load/unload functions
Sean Christopherson [Fri, 5 Mar 2021 01:10:59 +0000 (17:10 -0800)]
KVM: x86/mmu: Unexport MMU load/unload functions

Unexport the MMU load and unload helpers now that they are no longer
used (incorrectly) in vendor code.

Opportunistically move the kvm_mmu_sync_roots() declaration into mmu.h,
it should not be exposed to vendor code.

No functional change intended.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210305011101.3597423-16-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86: Defer the MMU unload to the normal path on an global INVPCID
Sean Christopherson [Fri, 5 Mar 2021 01:10:58 +0000 (17:10 -0800)]
KVM: x86: Defer the MMU unload to the normal path on an global INVPCID

Defer unloading the MMU after a INVPCID until the instruction emulation
has completed, i.e. until after RIP has been updated.

On VMX, this is a benign bug as VMX doesn't touch the MMU when skipping
an emulated instruction.  However, on SVM, if nrip is disabled, the
emulator is used to skip an instruction, which would lead to fireworks
if the emulator were invoked without a valid MMU.

Fixes: eb4b248e152d ("kvm: vmx: Support INVPCID in shadow paging mode")
Cc: stable@vger.kernel.org
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210305011101.3597423-15-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: nVMX: Defer the MMU reload to the normal path on an EPTP switch
Sean Christopherson [Fri, 5 Mar 2021 01:10:57 +0000 (17:10 -0800)]
KVM: nVMX: Defer the MMU reload to the normal path on an EPTP switch

Defer reloading the MMU after a EPTP successful EPTP switch.  The VMFUNC
instruction itself is executed in the previous EPTP context, any side
effects, e.g. updating RIP, should occur in the old context.  Practically
speaking, this bug is benign as VMX doesn't touch the MMU when skipping
an emulated instruction, nor does queuing a single-step #DB.  No other
post-switch side effects exist.

Fixes: 41ab93727467 ("KVM: nVMX: Emulate EPTP switching for the L1 hypervisor")
Cc: stable@vger.kernel.org
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210305011101.3597423-14-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Set the C-bit in the PDPTRs and LM pseudo-PDPTRs
Sean Christopherson [Fri, 5 Mar 2021 01:10:54 +0000 (17:10 -0800)]
KVM: x86/mmu: Set the C-bit in the PDPTRs and LM pseudo-PDPTRs

Set the C-bit in SPTEs that are set outside of the normal MMU flows,
specifically the PDPDTRs and the handful of special cased "LM root"
entries, all of which are shadow paging only.

Note, the direct-mapped-root PDPTR handling is needed for the scenario
where paging is disabled in the guest, in which case KVM uses a direct
mapped MMU even though TDP is disabled.

Fixes: d0ec49d4de90 ("kvm/x86/svm: Support Secure Memory Encryption within KVM")
Cc: stable@vger.kernel.org
Cc: Brijesh Singh <brijesh.singh@amd.com>
Cc: Tom Lendacky <thomas.lendacky@amd.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210305011101.3597423-11-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Fix and unconditionally enable WARNs to detect PAE leaks
Sean Christopherson [Fri, 5 Mar 2021 01:10:52 +0000 (17:10 -0800)]
KVM: x86/mmu: Fix and unconditionally enable WARNs to detect PAE leaks

Exempt NULL PAE roots from the check to detect leaks, since
kvm_mmu_free_roots() doesn't set them back to INVALID_PAGE.  Stop hiding
the WARNs to detect PAE root leaks behind MMU_WARN_ON, the hidden WARNs
obviously didn't do their job given the hilarious number of bugs that
could lead to PAE roots being leaked, not to mention the above false
positive.

Opportunistically delete a warning on root_hpa being valid, there's
nothing special about 4/5-level shadow pages that warrants a WARN.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210305011101.3597423-9-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Check PDPTRs before allocating PAE roots
Sean Christopherson [Fri, 5 Mar 2021 01:10:51 +0000 (17:10 -0800)]
KVM: x86/mmu: Check PDPTRs before allocating PAE roots

Check the validity of the PDPTRs before allocating any of the PAE roots,
otherwise a bad PDPTR will cause KVM to leak any previously allocated
roots.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210305011101.3597423-8-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Ensure MMU pages are available when allocating roots
Sean Christopherson [Fri, 5 Mar 2021 01:10:50 +0000 (17:10 -0800)]
KVM: x86/mmu: Ensure MMU pages are available when allocating roots

Hold the mmu_lock for write for the entire duration of allocating and
initializing an MMU's roots.  This ensures there are MMU pages available
and thus prevents root allocations from failing.  That in turn fixes a
bug where KVM would fail to free valid PAE roots if a one of the later
roots failed to allocate.

Add a comment to make_mmu_pages_available() to call out that the limit
is a soft limit, e.g. KVM will temporarily exceed the threshold if a
page fault allocates multiple shadow pages and there was only one page
"available".

Note, KVM _still_ leaks the PAE roots if the guest PDPTR checks fail.
This will be addressed in a future commit.

Cc: Ben Gardon <bgardon@google.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210305011101.3597423-7-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Allocate pae_root and lm_root pages in dedicated helper
Sean Christopherson [Fri, 5 Mar 2021 01:10:49 +0000 (17:10 -0800)]
KVM: x86/mmu: Allocate pae_root and lm_root pages in dedicated helper

Move the on-demand allocation of the pae_root and lm_root pages, used by
nested NPT for 32-bit L1s, into a separate helper.  This will allow a
future patch to hold mmu_lock while allocating the non-special roots so
that make_mmu_pages_available() can be checked once at the start of root
allocation, and thus avoid having to deal with failure in the middle of
root allocation.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210305011101.3597423-6-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
3 years agoKVM: x86/mmu: Allocate the lm_root before allocating PAE roots
Sean Christopherson [Fri, 5 Mar 2021 01:10:48 +0000 (17:10 -0800)]
KVM: x86/mmu: Allocate the lm_root before allocating PAE roots

Allocate lm_root before the PAE roots so that the PAE roots aren't
leaked if the memory allocation for the lm_root happens to fail.

Note, KVM can still leak PAE roots if mmu_check_root() fails on a guest's
PDPTR, or if mmu_alloc_root() fails due to MMU pages not being available.
Those issues will be fixed in future commits.

Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210305011101.3597423-5-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>